I am trying to develop a clinical risk score using a prediction model to predict the risk of developing a severe outcome (a type of tumour) that usually occurs in older adults from less severe clinical manifestations of the disease usually occurring in children (all chosen based on clinical expert knowledge on the disease).

The data I'm using comes from an ongoing clinical database using both prospective and retrospective data collection (If patient x joins at 50 years old but was diagnosed with clinical manifestation a at 25 years old, a will be entered in the data with the date at which it was diagnosed). All individuals in it have the same genetic disease. Participants joined the study at varying ages, some as babies and some over 60 years old.

The variables my data contains include the date of birth for every participant, their age when they joined the study and the date of their last follow-up. As for the clinical manifestations of the disease, they are all binary variables (Present/absent), while the one used as my outcome also comes with a second variable stating the age at diagnosis.

I'm having trouble defining the end of follow-up in my data. Having the birthdate of participants and knowing that outcome never happens before the age of 10, I'm thinking of choosing either birth or age of 10 as a starting point.

I'm struggling with the end of follow-up since it's an ongoing database, meaning that there is no date at which the study ends.

One possibility is just to use either date at the last visit for the censor and the date of outcome diagnosis for the event. Knowing that the oldest person diagnosed with the outcome is 69 years old, but that several people not developing the outcome are followed well until their eighties, does this pose a problem? Is there supposed to be a specific cutoff usually for the end of follow-up, or is it just what we're used to seeing for studies with a specific end time?

I'm confused since, in this case, we're not interested in something like remission after 5 years; we are just interested in knowing how certain predictors' presence affects the risk of developing the outcome.

As for my question, is this a situation in which Cox, or maybe even survival analysis, is maybe not the best answer?

If Cox regression is a good choice, how is follow-up usually defined in a situation like this?


1 Answer 1


Is there supposed to be a specific cutoff usually for the end of follow-up, or is it just what we're used to seeing for studies with a specific end time?

There is no need to have "a specific end time." An advantage of survival analysis is that you can use all the information that you have with respect to outcome times, even those with very long observation times without events. If you are using age as your time scale and the last event happened at age 69, then participants observed past age 69 will provide information up to age 69 in the Cox model. There's no reason (or advantage) to censoring those values at ages lower than what you have observed. As more data become available and events are found at later ages, data from those individuals will help refine the model.

As discussed in a previous question you do need to treat the study-entry age as left truncated for those who enter without having experienced the event, as no such individual provides information about potential events that might have occurred prior to the age at study entry.

Some cases

If patient x joins at 50 years old but was diagnosed with clinical manifestation a at 25 years old, a will be entered in the data with the date at which it was diagnosed

might need special treatment. For them, you might only have an upper limit to the age at which the event of interest occurred; it might have been earlier than 25 years old in that example. Those event ages should be treated as left-censored unless the specific age at the event is known. The Klein and Moeschberger text provides many examples of how to distinguish the different types of censoring and truncation.

  • $\begingroup$ Thank you for answering another of my questions! As for the left truncation, with the way the data was collected, isn't it the same as if I'm analyzing data as if every participant was followed since birth? For example, is there a difference in how I should analyze a participant who entered the study at 25 and developed the outcome at 40 vs. a participant who entered the study at 50 and also develop the outcome at 40? I have a problem seeing the difference in both since, in each case, if birth is chosen as the start, both will have developed the outcome after 40 years of follow-up. $\endgroup$
    – floubert
    Commented Sep 15, 2022 at 15:24
  • $\begingroup$ @floubert if time = 0 is birth (or 10 years of age), the participant who entered the study at 25 provides no information about event occurrence before 25. Therneau and Grambsch put it simply on page 75: "the patient was not at risk for an observable [event prior to study entry]...Such data, where the patient enters the risk set after time 0, is said to be left truncated." It's not about the patient's own event, it's about what other event times she should be included in for comparison. The counting-process data format handles that directly. $\endgroup$
    – EdM
    Commented Sep 15, 2022 at 16:34
  • 1
    $\begingroup$ @floubert such an individual doesn't represent individuals who might have had an event prior to study entry at age 25 and didn't enroll at all. The contribution to the survival curve thus only starts at age 25: the event at age 40 was observed conditional on already having been event-free through age 25. The Therneau and Grambsch example is such a case: date of diagnosis was prior to study entry and data were available from prior to study entry, but left truncation at study entry was needed to account for that entry conditional on a prior event-free period. $\endgroup$
    – EdM
    Commented Sep 15, 2022 at 20:04
  • 1
    $\begingroup$ @floubert this study is tricky: do you have a representative sample of all those with this genetic defect, or are you perhaps biased in favor of those who experience the serious later event of interest? Even when left truncation is included there might be trouble, as standard analysis assumes that event times and (left-truncated) delayed-entry times are independent. I suspect that some accrual into your study is driven by having that event, which would invalidate that assumption. This probably requires consultation with a highly experienced survival analysis expert (which I am not). $\endgroup$
    – EdM
    Commented Sep 15, 2022 at 20:11
  • 1
    $\begingroup$ @floubert this answer illustrates how it's handled. In the Surv(startTime,stopTime,event) counting-process data format, the startTime is a left-truncation time (age at entry in your case). A data row simply isn't used in the Cox calculations until its own startTime. That handles delayed entry as in your case, and other things like time-varying covariates. See Section 3.7 of Therneau and Grambsch. $\endgroup$
    – EdM
    Commented Sep 16, 2022 at 14:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.